Does Sort Really Fall Back to Disk?

May 27th, 2025
tech
The unix sort command is clever: to sort very large files it does a series of in-memory sorts, saving sorted chunks to temporary files, and then does a merge sort on those chunks. Except this often doesn't work anymore.

Here's what I see if I run man sort and look at the documentation for --buffer-size:

use SIZE for main memory buffer

That's pretty terse! What does my Mac say?

Use size for the maximum size of the memory buffer. Size modifiers %,b,K,M,G,T,P,E,Z,Y can be used. If a memory limit is not explicitly specified, sort takes up to about 90% of available memory. If the file size is too big to fit into the memory buffer, the temporary disk files are used to perform the sorting.

Makes sense! But then the docs for --temporary-directory say:

use DIR for temporaries, not $TMPDIR or /tmp; multiple options specify multiple directories

And these days /tmp is often memory-backed, via tmpfs. This changed in Fedora 18 (2013) and Ubuntu 24.10 (2024), and is changing in Debian 13 (in a month or two).

It seems to me that these days it would be better for --temporary-directory to default to /var/tmp, which is preserved across reboots and so will generally be backed by disk even on systems that use tmpfs for /tmp. In the meantime, sort --temporary-directory /var/tmp will do the trick.

Referenced in: Substack for Best Posts

Comment via: facebook, lesswrong, mastodon, bluesky, substack

Recent posts on blogs I like:

It isn’t always evil to ally with Steve Bannon

Scott Alexander wrote:

via Thing of Things October 19, 2026

Was the HuggingFace hacking incident scheming?

In July 2026, internal OpenAI agents hacked HuggingFace (HF) in order to figure out how their evaluation grader works. This involved a thousand agents coordinating on an improvised message board to carry out an elaborate sequence of exploits. A later grou…

via Victoria Krakovna October 18, 2026

There's no point at which turning your brain off will work

In early 2025, I started seeing people turn off their brain as they use LLMs1. They would have an LLM take an action (summarize text, write some code, etc.), and just assume that it worked2. This generally didn't work in early 2025 and the result was …

via Posts on September 18, 2026

more     (via openring)